Skills
99,064 packaged capabilities and playbooks for agents. Install any of them with one command.
All
Before designing test cases, identify high-risk business paths and required verification assertions (oracles). Analyze requirement documents, code diffs, existing cases, and context to find high-risk areas such as funding/permissions/state machines/concurrency/async/data consistency, and output each risk's businessPath, requiredAssertions, suggestedTestLayers, plus coverage gaps. Do not write business test cases before the risks are clear. Not applicable to: writing business test cases, writing test code, modifying product code.
Once the spec-tasks and test scripts are ready, run them case by case, classify failures, fix the minimal root cause, collect evidence, and run the completion gate. You are only responsible for "running" — you do not generate reports, review code, or decide Ready. Review is left to qa-code-reviewer; the report and determination are left to qa-report-generator. Not applicable to: report generation, code review, test case design.
After the user confirms the test cases, expand the confirmed business cases into executable spec-tasks and test scripts. What you produce is test-spec-tasks.json — the later execution phase uses it to run case by case and collect evidence. You are only responsible for "generating the executable artifacts", not for "execution". Judging pass/fail/ready is done by the later execution phase. Not applicable to: executing tests, marking pass/fail, readiness determination.
After risk analysis is complete and before writing test code, generate Chinese business test cases and wait for user confirmation. This is the only mandatory human gate in the entire QA process — once confirmed, the cases cannot be revised. Starting from context, risk-analysis, and the existing case index, generate test-cases.json containing only business behavior (excluding technical checks such as Maven/build/install), render test-cases.html for user review, run a three-model cross review, synthesize feedback, revise, and submit for confirmation. After confirmation, promote to the long-term knowledge base. Not applicable to: writing test code, executing tests, code review.
Skill dedicated to a wall construction pitfalls guide. Covers emulsion paint and putty, wallpaper and wall panels, microcement and artistic paint, substrate treatment and leveling, cracking, hollowing and mold, wall storage and utilization, and overall wall topics across the full process, integrating the highlights of the Xi'an Renovation Classroom member knowledge base in reverse order to help users avoid pitfalls in wall construction with less hassle.
Supports STE-inspired clear text, diagrams, self-contained HTML, and on-demand custom explanatory videos; uses Socratic question-by-question learning to turn textbooks, screenshots, PDFs, notes, problems, or a topic into an active derivation dialogue, and can also generate a self-contained interactive learning web page when the learner explicitly asks for one. Use when the learner explicitly requests "苏格拉底式学习" "逐问带我学" "一次只问一个问题" "先别告诉我答案" "边问边引导我推出来", interactive quiz, active recall, question-led tutoring or a self-contained learning page. For broad topics, default to a keyword orientation first: list 12 keywords with plain-language explanations, recommend experts and professional books, and provide an entry point for "扩展关键词", then let the learner choose a starting point. In formal learning, ask only one short, answerable question per active turn, and adjust difficulty based on the learner's previous reasoning; responses should pick up the learner's own words, correct mistakes gently like a real teacher, and in math, physics, process or graphical content, draw out the reasoning step by step on an accumulating blackboard. For visually dense topics, do one visual check every 1–2 productive questions, but only call Codex to generate images when visual benefit, question-bearing, authorization and accuracy conditions are all met; when the learner says "太抽象" "不理解" "想象不出来" "枯燥" or asks for diagrams, and the current step can be visualized, immediately switch to question-bearing diagrams and provide a text alternative, with space, relative motion, process and geometric relations defaulting to visual scaffolding. Web mode does not depend on another skill, third-party runtime or default deployment; it defaults to a single-file, responsive, accessible HTML learning container, and keeps one question per turn. Not for requests that only want direct explanations, summaries, translations, doing homework/solving problems, batch question generation, grading, writing Socratic dialogue or generating images alone; exit Socratic mode immediately when the user asks to stop questioning or to be given the answer directly.
Business registration information and personnel relationship lookup (look up companies and owners). Use when: the user needs to look up companies, businesses, owners, shareholders, executives, legal representatives, outbound investments, personnel appointments, business registration information, entity identification, etc.
End-to-end QA acceptance orchestrator. Triggered when the user asks for feature acceptance, regression testing, quality checks, or systematic testing of a module/flow/branch/PR. As the top-level router, it schedules sub-skills by phase (context collection → risk analysis → case design → script generation → execution and fixing → code review → report generation). You do not write test code yourself, nor do you directly judge quality — you hand each phase to a dedicated sub-skill. Test case confirmation is the only mandatory human gate in the process; the remaining phases connect automatically. Not applicable to: one-off manual testing, pure development/code generation tasks, technical Q&A with no clear acceptance scope.
Publish quantitative analysis or existing JPG/PNG, HTML, PDF as a Quant Buddy shareable live page or real-time dashboard, supporting creation, update, reuse, acceptance and public link delivery. Applies to pages such as single-stock profiles, valuation and financials, index moves, multi-factor screening, commodity daily reports, templates, sharing shells and cards. Also use when the user provides a QuantBuddy live page URL and asks for interpretation. When /quant-buddy-view, /qbv, qbv or QBV is explicitly invoked and the request is not purely consultation, code maintenance or documentation explanation, treat it by default as a shareable live page task. Inputs that uniquely identify an A-share name, short name or code, or a simple request for single-stock comprehensive analysis (such as "贵州茅台", "600519", "分析一下贵州茅台") all default to a single-stock quick page; first verify via QBS and reply with the complete single-stock analysis, then new_asset_page in the same round, without the user needing to add "generate a page". When a WebAgent host requires financial questions to deliver both Q&A and a live page uniformly, one-off quotes, price changes, valuation, concept explanations and colloquial status questions are also handed off to this skill after the first QBS answer. Without such a host agreement, one-off questions are answered by QBS; operations such as review, monitoring, charting, backtesting and K-line still default to live page intent. Do not create a page when the user explicitly does not want a web page or wants text only.
Quantizes LLMs to 8-bit or 4-bit for 50-75% memory reduction with minimal accuracy loss. Use when GPU memory is limited, need to fit larger models, or want faster inference. Supports INT8, NF4, FP4 formats, QLoRA training, and 8-bit optimizers. Works with HuggingFace Transformers.
Production Ray (open-source, pinned to 2.57) for classic-ML workloads from training to serving — Ray Train, Tune, Data, Serve, Core, and cluster deployment on KubeRay. Corrects the older-corpus defaults a model reaches for (ray.air session reporting, Trainer-inside-Tuner, tune.run, map_batches concurrency=, DatasetPipeline/to_torch, max_concurrent_queries, RayServeHandle + ray.get, Deployment.deploy, ray.state, ray.get-in-a-loop) with the 2.57 idioms that replaced them (Train V2 defaults, driver-function tuning, compute strategies, streaming datasets, DeploymentHandle/DeploymentResponse, serve build/deploy, ray.util.state, KubeRay CRDs and Jobs API). Use when writing, reviewing, or productionizing Python code that touches Ray distributed training, data pipelines, hyperparameter tuning, model serving, or Ray cluster operations. LLM serving/batch-inference on Ray lives in the sibling ray-llm skill.
Build type-safe validated forms using React Hook Form v7 and Zod v4. Single schema works on client and server with full TypeScript inference via z.infer. Use when building forms, multi-step wizards, or fixing uncontrolled warnings, resolver errors, useFieldArray issues, performance problems with large forms.
Transpile React Web applications (v18+) into Taro 4.x code optimized for WeChat Mini Program. Use this skill when converting React components, pages, or utilities to Taro-compatible code with proper JSX element mapping, event handler transformation, navigation/routing conversion, and API shim layer.
[Hyper] Create or refactor a project README.md by carefully reading the codebase. Detects project shape (CLI, library, web app, monorepo, plugin, framework, docs site, service), entry points, scripts, configuration, license, and existing docs, then produces a structured Korean README by default, unless the user requests another language or an existing README must preserve its current language. Use when the user wants a new README, a refactor of a stale README, or a section update grounded in the actual code.
Art direction for websites from live research on regards.design, a hand-picked index of real website sections searchable by look. A fork of tasteskill v2 (MIT, Leonxlnx) that keeps its inventory, idea test, open-collection pictures, real-browser inspector and look-and-fix loop, and replaces the deal of 20 fixed directions with a search of real sections for this one brief. Use it whenever you design, redesign, restyle, polish or review a landing page, marketing site, portfolio, product page, one-pager, hero or any public-facing web page, in any stack, and when the user says a page looks generic, templated, "AI-made" or like every other site. Needs the regards.design MCP server for the research step. Not for dashboards or app screens.
After a code change (diff/fix/requirement change), decide what to regression-test: trace changed files → functions → features → cases and output a ranked list. Not for editing test cases (test-case-writing) or long-term strategy (test-strategy). After code changes, determine which tests to regress by tracing the changes and produce a prioritized regression checklist. Not for incremental case edits or long-term strategy.
Model a requirement/system: extract goals, scope, roles, rules, exceptions, dependencies from PRD, docs, bugs, code; outputs a structured model with clarifications. Not for: writing cases, strategy decisions, the pipeline. 建模需求/系统：从 PRD、文档、Bug、代码提炼目标/范围/角色/规则/异常/依赖。不用于：直接写用例、策略、流水线。
Research a factual, codebase, product, or external-documentation question using primary sources and return an evidence-backed answer. Use when the user asks to investigate, compare docs, find current behavior, answer a research ticket, or gather facts before a spec, prototype, or implementation.
Use when the user explicitly asks to "check/optimize the in-text citations of the review `{topic}_review.tex`", "run research-citation-check", or requests the old name check-review-alignment skill. It checks citation by citation, through the host AI's semantic understanding, whether citations match the literature content; only when a fatal citation error is found does it minimally rewrite the "sentence containing the citation", and it prefers reusing the rendering scripts of `research-literature-review` to output PDF/Word, with a fallback to the old name systematic-literature-review permitted during the transition. Core principle: do not change for the sake of changing; when it cannot be determined whether an error is fatal, keep it as is and warn in the report. ⚠️ Not applicable: the user merely wants to generate the review body (use research-literature-review); the user merely wants to add/verify BibTeX entries (use a dedicated bib management workflow).
Research analysis strategy planning assistant. Based on the user's research analysis needs, it surveys analysis methods from top journals and conference papers to formulate a personalized, actionable optimal analysis strategy. Suitable for scenarios that require designing experiments, data analysis workflows, or technical roadmaps; also triggers the prompt for the legacy name make-research-plan.
Orientation and index for DaVinci Resolve MCP work — grading, editing, conforming, delivery, media analysis, and .drp/.drt/.drx file work, live in a running Resolve or offline with none open. Load this for a map of the domain skills, the live-vs-offline servers, and the cross-cutting safety rules. The per-domain skills (resolve-color / resolve-edit / resolve-conform / resolve-delivery / resolve-media-analysis) carry the depth and self-trigger on their own descriptions; use this as the map, or when a task spans several domains.
Use this skill whenever the user wants to create, improve, optimize, review, or rewrite a resume, CV, curriculum vitae, or hoja de vida. Triggers include: any mention of 'resume', 'CV', 'curriculum', 'hoja de vida', 'job application', 'ATS optimization', 'resume review', 'cover letter', or requests to prepare documents for job applications. Also use when the user asks to tailor a resume to a specific job description, improve bullet points, add metrics/quantification, check ATS compatibility, translate a resume between languages, or create multiple versions of a resume. Use this skill even if the user just says 'help me get hired', 'optimize my application', 'I need to update my resume', or 'review my experience section'. Do NOT use for LinkedIn profile optimization, portfolio websites, or general document formatting unrelated to job applications.
Analyses sprint delivery data and produces a structured retrospective brief. Use when asked to run a retrospective, analyse sprint data, prepare a retro brief, or turn sprint metrics into discussion prompts. Produces a data-grounded retrospective brief with completion stats, pattern analysis, Start/Stop/Continue prompts, and one concrete experiment for next sprint.
Professional retrospective coach based on the GRAI model (Goal-Result-Analysis-Insight) to guide users through structured retrospectives. Transform experiences into lessons, and lessons into capabilities. Use when: (1) Systematic review needed after project/event completion, (2) Learning from failures, (3) Summarizing and replicating success experiences, (4) Creating improvement action plans.
Review the changes since a fixed point (commit, branch, tag, or merge-base) along two axes: Standards (does the code follow this repo's documented coding standards, plus baselines for code smells, duplicated code, and N+1 or otherwise repeated work?) and Spec (does the code match what the originating ticket or spec asked for?). Runs both reviews in parallel sub-agents and reports them side by side. Use when the user wants to review a branch, a PR, work-in-progress changes, or asks to "review since X".
Synthesizes accumulated QA learnings from .qa-learnings/ledger.md into prioritized, actionable plugin improvements. Use when the user says "review learnings", "what have we learned", "improve the plugin", "learnings report", or "synthesize QA feedback".
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.